Large language models frequently generate fluent explanatory narratives even when the un-derlying evidentiary basis for a claim is absent or indeterminate. This paper identifies a struc-tural failure mode responsible for this behavior: epistemic compression, in which uncertainty,missing primary data, and conditional inference are collapsed into authoritative narrative ex-planations.Unlike classical hallucination, epistemic compression does not require fabrication of falsefacts. Instead the failure occurs at the level of epistemic status: statements that should bepresented as uncertain or unsupported are expressed as if evidential confirmation exists.This work introduces four contributions:1. A taxonomy of epistemic failure modes in language models.2. The Evidence-First Protocol (EFP), a structured reasoning protocol.3. The Lambda Scan, a deterministic auditing procedure for evaluating claims.4. The Epistemic Compression Benchmark (ECB-100), a dataset designed to measureepistemic compression.Together these components provide a practical framework for detecting and mitigating epis-temic compression in language models